A Method of Ore Image Segmentation Based on Deep Learning

Abstract

In the analysis of ore particle size based on images, ore segmentation is a key link. After accurate segmentation, the geometric parameters such as the contour of these blocks, the external rectangle, the center of mass and the invariant moment can be further obtained, and the ideal ore particle size can be obtained effectively. A method of ore image segmentation based on deep learning is proposed in this paper. The method focuses on solving the problem of inaccuracy caused by mutual adhesion and shadow on the ore image. Firstly, complex environment image data set is obtained by using high resolution webcam; Next, we use the annotation data set to train HED (Holistically - Nested Edge Detection) model. This model can extract the image edge feature with strong robustness. Then, thinning edge is extracted using table lookup algorithm. The final step is labeling the connected region and getting segmented results. Our method is compared with the Watershed method based on gradient correction, and experimental results show the effectiveness and superiority of the proposed method.